Built-In Office AI: Is the License Worth It?

Today's list prices

Every figure is a list price as of 12 September 2026, per user, per month, in US dollars.

Two small things that can cost a lot:

• Paying monthly adds 14–20% on top of the annual commitment.
• There are seat limits, but they run the other way. One provider publishes no minimum; at the other, the three business plans can only be bought up to 300 users, and above that only the enterprise tier remains. The business plans of the separately subscribed assistants usually come with both a lower and an upper headcount limit.

What changed since 2025

Two moves explain why today's prices look different from the ones in most blog posts.

In January 2025 one provider dropped its separate AI add-on and folded the capabilities into the business base plans. The example in their own announcement: someone who had been paying $32 for a base plan plus the AI add-on paid $14 from then on — while the base plan itself went up by two dollars.

The date matters, because plenty of sources get it wrong: this was a 15 January 2025 announcement, effective March 2025. Several articles put the same change in March 2026 with different figures — that is incorrect.

The other provider moved the opposite way: it kept the $30 enterprise add-on, and on 2 December 2025 a tier aimed at smaller companies arrived next to it at $21, which today shows as $18. So the entry point got 40% cheaper while the enterprise price did not budge.

(Whether the $18 is still a promotion is unclear: the announced discount ran to June 2026, yet the price list keeps showing it. Worth clarifying before you buy.)

The question is not whether there is AI. It is how much

This is the most important change of 2026, and most comparisons skip it.

The AI is in the plan — but with a quota. One provider's public table shows how much more the separately purchasable expanded access gives you:

On the entry tier there is on top of that no AI at all in documents, spreadsheets, slides or video calls — only in the standalone assistant, with three audio summaries a day.

The tightest cap is image generation: thirty images a month. For a marketer that is half a day. The hundred-slide monthly budget behaves the same way — and what is worth handing to a machine there at all is measured in the from outline to deck post, with numbers.

The same pattern turned up in project tools: AI became part of the paid plan there too, but usage moved onto credits — the project and task management post goes into that in detail.

Payback: five measurements and one formula

Payback can be calculated, and the result is not what most people expect.

The formula:

Monthly license fee (USD/person)
────────────────────────────────  =  Hours to save per month to break even
Fully loaded hourly cost (USD/h)

What the measurements give (on a 168-hour month):

And break-even (our own calculation with the formula above):

The conclusion: even the control-group measurement (4.7 hours) is roughly 1.6 to 13 times what break-even needs, at every pay level above. So the license fee is not the bottleneck.

The risk sits elsewhere, and there are numbers for that too:

• They do not use it. Daily or weekly use ranges from 33% to 80% across the measurements, and in one public-administration trial, 46% of the non-users named privacy and security concerns as their main reason — not that they could not work the tool.
• Checking takes it back. In another trial 60% of the participants made moderate or substantial edits to the output, and up to 7% of them ended up with more work because of it.
• And the hardest counterpoint: in the only randomised, stopwatch-measured study, experienced professionals were 19% slower — while they felt 20% faster.

So the question is not whether it is worth buying. It is whether people will actually start using it, and whether your work sits in the part where it measurably helps — the where AI saves real time post breaks that down task type by task type.

Data handling: same model, different rules

This section matters more than the price, and plenty of people have it exactly backwards.

What decides is not "which AI" you use, but which subscription you use it on. The same model falls under different data handling on the free, consumer and business tiers. A March 2026 professional guide puts it precisely: whether your prompt is later used for training is decided not by the model's architecture but by the provider's contract terms, technical configuration and policy.

On the business plans the statement is uniform: they do not train on your content. On the consumer tiers, by contrast, some providers state explicitly that content is used to train the models, with an opt-out.

So the difference is not whether they train, but:

• Retention period. At one provider it varies by product: from 90 days to indefinitely for the office integration, at the administrator's discretion; up to 36 months for the standalone assistant; nothing after the session for a third product.
• Data residency. For European users one provider publishes separate guarantees about the EU data boundary: EU traffic stays inside the boundary, while worldwide traffic can be routed to the EU and elsewhere for processing.
• Administrator control. Who can switch it on, what they get to see, how long it is stored.

The practical rule: if someone at the company works on company material through a free or personal account, that is not "the same thing, cheaper". It is a different contract.

Beyond English: here the difference is real

This is the most concrete language angle in the whole decision.

In other words: one plan also drafts in your language, the other mainly summarises and analyses in it. If most of the daily work is writing email and documents in a language other than English, that row weighs more than a few dollars a month.

One caveat on both: the documentation itself says the assistant supports fewer languages than the interface, and neither provider publishes per-language quality data.

Common misconceptions

• "Expensive, not worth it." The arithmetic says otherwise. Break-even needs under an hour and a half a month at most pay levels, and the control-group measurement shows 4.7 hours.
• "If it is in the plan, it is unlimited." It is not. The quotas are public, and on the lower tiers they are tight — thirty images a month, a hundred slides a month.
• "Everyone will use it." The measurements show daily or weekly use between 33% and 80%, and the biggest obstacle is uncertainty about privacy, not usability.
• "AI is AI, the subscription makes no difference." It makes a difference. Retention period, data residency and the training statement vary by subscription — for the very same model.
• "Both do the same in every language." No. Drafting help works beyond English in one of them; in the other it is capped at eight languages, and most languages are not among them.

What to take away

Three questions, in this order, and price is only the third.

First: does it have to draft in your own language? If yes, that settles most of the decision.

Second: who will use it, and what do they spend the day on? If the day is writing, closing documents and answering customers, the measurements show a gain. If it is meetings and judgement calls, they do not.

Third: is the quota enough? Look at the monthly budget for image generation, slide generation and the spreadsheet function before you look at the monthly fee. In 2026 that is the real limit.

And a fourth one that is not measured in money: which subscription the company's data ends up on. That is not a setting, it is a contract.